Sentiment Analysis
- Tags
- text-processing
Extract [see page 8, emotions, sentiments or opinions] from texts given by humans.
SA is used to make decisions, answering questions such as:
SA can be seen as equivalent to Opinion Mining, even though a sentiment does not necessarily express an opinion.
Segmentation often involves breaking an input text into:
| Term | Meaning | Example |
|---|---|---|
| [see page 18, Target of Opinions] | A person/event/organisation. Represented as a hierarchy of features (phone->battery->life). | This is an iPhone Battery. |
| Opinion | Something expressed about a feature. | My iPhone's Battery is too short. |
| Objective | Refers to the data in the text. | I payed £500.00 for this. I got a refund. |
| Holder of Opinions | Who this text is written by? | I'm me, not you, me :-) |
Granularity
Document Level
Find the [see page 27, opinion] expressed in a complete document.
Is the document in overall expressing a positive or negative opinion (eg. Movie review).
Sentence Level
Consists of 2 steps:
Feature Level
[see page 31, Provides] more specific opinions on specific features of the object. The goal is provide a fine grained analysis of sentences (which can contain both +ve and -ve opinions).
For example do people like the screen of this phone, but not the camera?
Runs in 5 steps:
- Identify entities or objects.
- Extract object features that have been commented on by the opinion holder.
- Group similair features (eg. screen and touch screen, power-usage and battery life)
- Classify the opinions as +ve, -ve, neutral.
- Produce a summary of all feature based opinions.
Challenges
- [see page 44, identifying] all components of an opinion and their relations. Having two entities with different opinions scattered around the text is difficult to follow.
- [see page 45, synonym] match. "Voice == Sound Quality".